[Inductor] add inductor config: masked_vec (pytorch#134566)

This PR adds inductor config: masked_vec to control enable/disable masked vectorization for the tail_loop, and enable by default.

Pull Request resolved: pytorch#134566
Approved by: https://github.com/jgong5, https://github.com/jansel

torch/_inductor/codegen/cpp.py

@@ -3778,7 +3778,11 @@ def run(kernel):
                 )
                 main_loop.set_kernel(vec_kernel)
                 main_loop.simd_vec = True
-                if could_masked_vec and (tail_loop.size - tail_loop.offset) >= 4:
+                if (
+                    config.cpp.enable_loop_tail_vec
+                    and could_masked_vec
+                    and (tail_loop.size - tail_loop.offset) >= 4
+                ):
                     tail_loop.steps = tail_loop.size - tail_loop.offset
                     masked_vec_kernel = codegen_kernel(
                         CppVecKernel,
@@ -3818,7 +3822,7 @@ def run(kernel):
                 )
                 inner_main_loop.set_kernel(tile2d_kernel)
 
-                if could_masked_vec:
+                if config.cpp.enable_loop_tail_vec and could_masked_vec:
                     (
                         inner_main_loop_of_outer_tail_loop,
                         inner_tail_loop_of_outer_tail_loop,

torch/_inductor/config.py

@@ -768,6 +768,9 @@ class cpp:
     # decomposed into 7x4x2 thread blocks along MxNxK of a GEMM.
     gemm_thread_factors = os.environ.get("TORCHINDUCTOR_CPP_GEMM_THREAD_FACTORS", None)
 
+    # Whether to enable masked vectorization for the tail_loop.
+    enable_loop_tail_vec = True
+
 
 # config specific to codegen/triton.py
 class triton: